Research And Application Of Machine Learning In Trend Prediction

Satellite is a large and complex system which integrates many kinds of the most sophisticated technologies.Its role has penetrated into various fields and occupied an irreplaceable position.However,satellites are exposed to the harsh cosmic environment all the year round,and will be interfered by external factors such as solar activity,cosmic rays,and space debris and so on.At the same time,due to internal factors such as workload and aging of equipment,they will cause failures.In the process of long-term monitoring and management of satellite,a large amount of satellite telemetry data will be generated,from which various useful information can be extracted,and the potential change rules can be effectively used to predict future trends,which has a particularly important practical significance for the early detection of satellite anomalies.In this thesis,machine learning algorithm is applied to trend prediction of satellite key parameters.The main research contents are as follows:(1)Aiming at the problem of satellite key parameters affected by noise,data missing and machine learning prediction model selection,several common methods and three evaluation indexes are analyzed and compared.An adaptive wavelet threshold denoising algorithm based on BAS optimization is proposed.The experimental data show that the algorithm is feasible and effective.(2)Trend prediction is implemented using fast learning network(FLN),Elman neural network,and echo state network(ESN)in a shallow learning model.There is no direct relationship between the current state value and the previous state value of traditional ESN.Therefore,the state of neurons at the previous time can be controlled by adding adjusting parameters ? to improve the memory ability of ESN.A technique is proposed to improve the properties of ESN solution,which performs ridge regression in the reservoir state space instead of the linear regression,so that the amplitude of the output weight can be effectively adjusted.In the ESN training process,for the ESN parameter optimization problem,the salp swarm algorithm in the swarm intelligence optimization algorithm is used to optimize the important parameters of ESN,which makes the prediction model is more accurate(3)A combined prediction model based on modified ensemble empirical mode decomposition(MEEMD)and long and short term memory(LSTM)network is proposed to improve the prediction accuracy of single model for non-linear and non-stationary time series Firstly,the original sequence is decomposed into several intrinsic mode components(IMF)and residual components at different time scales by using MEEMD.Then,the LSTM neural network is used to predict each component.Finally,the prediction results of multiple components are fused to obtain the actual prediction results(4)Several machine learning models mentioned are applied to the lithium ion battery capacity trend prediction example to verify the validity of the model.The experimental results show that prediction accuracy of the echo state network in the shallow learning model is higher.The combination of the swarm intelligence algorithm can improve the prediction accuracy of the traditional ESN neural network.The MEEMD-LSTM model in the deep learning model has a better prediction effect(5)A comprehensive electronic system health monitoring software system is developed by using MATLAB and C#language.